---
language:
- ar
- ary
library_name: transformers
license: apache-2.0
base_model: ilyaslbern7347/whisper-darija-stage1
pipeline_tag: automatic-speech-recognition

tags:
- whisper
- automatic-speech-recognition
- speech-recognition
- darija
- moroccan-arabic
- whisper-finetuned
- generated_from_trainer

metrics:
- wer

model-index:
- name: whisper-darija-vols-stage2
  results:
  - task:
      type: automatic-speech-recognition
      name: Automatic Speech Recognition
    dataset:
      name: Moroccan Darija Flight Queries
      type: custom
    metrics:
    - type: wer
      value: 23.28
      name: Word Error Rate
---

# Whisper Darija Flights Stage 2

This model is a fine-tuned version of **ilyaslbern7347/whisper-darija-stage1** using a custom dataset of Moroccan Darija flight-related speech.

The objective of this model is to improve speech recognition accuracy for Moroccan Darija users searching for flights using natural language.

## Model Description

This model is based on OpenAI Whisper and has been further fine-tuned specifically for Moroccan Darija speech in the travel domain.

The model recognizes spoken requests such as:

- Book a flight
- Search for flights
- Departure city
- Destination city
- Travel date
- Number of passengers
- Flight-related conversational requests

This model is intended to be integrated into intelligent flight booking assistants.

---

# Base Model

- **Base Model:** ilyaslbern7347/whisper-darija-stage1

---

# Intended Uses

This model is suitable for:

- Automatic Speech Recognition (ASR)
- Moroccan Darija transcription
- Voice assistants
- Flight booking assistants
- Conversational AI
- Travel applications

---

# Limitations

This model was fine-tuned only on Moroccan Darija speech related to flight booking.

Performance may decrease for:

- General conversations
- Medical vocabulary
- Legal vocabulary
- Noisy audio
- Strong regional accents not represented in the training data

---

# Dataset

The model was trained on a custom Moroccan Darija speech dataset containing flight-related queries.

The dataset includes recordings covering:

- Departure cities
- Arrival cities
- Dates
- Passenger counts
- Flight reservations
- Flight search requests

---

# Training Procedure

## Hyperparameters

| Parameter | Value |
|-----------|-------|
| Learning Rate | 5e-6 |
| Train Batch Size | 8 |
| Eval Batch Size | 8 |
| Gradient Accumulation | 2 |
| Total Batch Size | 16 |
| Warmup Steps | 40 |
| Max Training Steps | 200 |
| Weight Decay | 0.01 |
| FP16 | True |
| Gradient Checkpointing | True |
| Optimizer | AdamW |
| LR Scheduler | Linear |

---

# Training Results

| Step | Training Loss | Validation Loss | WER |
|------|--------------:|----------------:|-----:|
| 50 | 1.5132 | 0.4115 | 30.17 |
| 100 | 0.1160 | 0.2923 | 25.51 |
| 150 | 0.0102 | **0.2779** | **23.28** |
| 200 | 0.0048 | 0.2775 | 23.46 |

Best checkpoint:

- Validation Loss: **0.2779**
- WER: **23.28**

---

# Evaluation

The model achieved:

- **Word Error Rate (WER): 23.28%**

This represents a significant improvement over the base model for Moroccan Darija flight-related speech recognition.

---

# Example Use Cases

The model can transcribe requests such as:

> "بغيت نحجز رحلة من كازا لباريس."

> "شنو أرخص رحلة لغدا؟"

> "بغيت نمشي لطنجة نهار الجمعة."

---

# Framework Versions

- Transformers
- PyTorch
- Datasets
- Tokenizers

---